I did some performance profiling of H::T some time ago -> it turns out 
that there is quite a bit of speed up (somewhere between 10% and 10x, 
depending on the page) by re-writing part of the code located around 
line 2660.

Does this pass all the .t tests? If it does then shouldn't this get 
accepted into H::T as a patch? Care to supply the change as a diff?

Sorry for the shameless blog plug, but I only learnt how to use diff and 
patch recently and blogged about how to to do:-
Not sure if it passes every test -> I use a heavily modified version of H::T which supports dynamic loading of escape-modules, support for arbitrary TMPL_xxx syntax, an odd bug fix + some other goodies... so I haven't specifically tested it against 2.10.   That same reason also makes it hard to generate an isolated patch.

I have posted this code before, but AFAICR there wasn't much discussion of merging it.